New trends and developments in automotive industry Part 2 pptx

New Trends and Developments in Automotive Industry

20




Fig. 5. Sting support system (left) and truck setup in the 9.5x9.5 test section (right)
LLF has two standard support systems (Fig. 4 and 5):
• A sting support system, to which models are mounted via an internal balance. This is
mostly used for aircraft and helicopter testing, but is also necessary for cars in
combination with the moving belt (moving ground plane)
• An external under-floor platform type balance with a model support by pads or special
adapters. This is mostly the primary choice for cars and trucks
3. Force and moment measurements
Two test sections of 20 m length each are suitable for automotive testing at DNW.
Full-scale cars and vans are often tested in the 8x6 closed test section, where up to 400 km/h
wind speed can be reached. A car platform with adjustable pads is linked to an external
underfloor balance for the measurement of six stationary load components, i.e., drag, side
force, lift, and the moments in roll, pitch and yaw. The platform can be used in combination
with a tunnel floor boundary layer control system which injects pressurized air into the
sublayer upstream of the test vehicle to improve the road simulation conditions and can be
used in both test sections.
A sting system for internal balance (six-component) supported cars can be used in
combination with a moving belt system for perfect road simulation and rolling wheels.
Full-scale trucks and buses use the 9.5x9.5 closed test section, where the maximum wind
speed of 200 km/h is more than enough. A truck platform is available in combination with
loose air cushion elements to support the large and heavy vehicles. This setup provides
three stationary load components, i.e. drag, side force and yawing moment by means of the
external underfloor balance.

The external balance (EXB) is a six-component platform balance, equipped with three
horizontal load cells with a resolution of 0.15 N and three vertical load cells with a
resolution of 0.30 N. This balance is installed underneath the test section. The test vehicle
rests with its wheels on four small pads which are flush with the floor of the turntable and
can be adjusted over a wide range to match track and wheelbase. The wheel pads are
incorporated in a rigid supporting frame (car platform), which is connected to the metric
part of the balance. When use is made of the 8x6 test section, the vertical distance between
supporting frame and balance is bridged by a spacer. The balance assembly can rotate over
± 180° in increments of ± 0.1°.
Automotive Testing in the German-Dutch Wind Tunnels

21

Fig. 6. Model on sting above moving belt


Fig. 7. Truck in 9.5x9.5 test section
4. Road simulation
During wind tunnel tests the relative motion between vehicle and road and the rotation of
the wheels is often disregarded. The road is then represented by the rigid floor of the test
section and the vehicle rests with stationary wheels on the pads of the balance platform.
The flow pattern around such a configuration is principally different from that on the road
due to the grown boundary layer along the wind tunnel floor. The boundary layer thickness
near the model may reach a thickness of half the ground clearance of a standard passenger
car. This will affect the aerodynamic phenomena around the car.
The effects from various ground simulation techniques at automotive testing in a wind
tunnel have been discussed in years around 1990 in various papers of the Society of
Automotive Engineers (SAE). Mercker and Knape (1989) discussed the ground simulation
with a moving belt or with tangential blowing. Mercker and Wiedemann (1990) compared
the results obtained at different ground simulation techniques.

New Trends and Developments in Automotive Industry

22

Fig. 8. Effects from ground clearance (left) and rotating wheels (right)
Another deficiency arises from the stationary wheels. Rolling wheels not only affect the flow
over and around the wheels but also the overall flow pattern around the car.
As long as the drag of the tested vehicle is relatively high, the error on the drag from the
fixed ground floor and the stationary wheels is generally negligible. For vehicles with lower
drag this error is not negligible, especially when the low drag is achieved by measures at the
underbody of the vehicle. These effects are illustrated in figure 8.
The results from wind tunnel tests can become more realistic when the effect from the
boundary layer development along the wind tunnel floor is reduced by boundary layer
suction or tangential blowing. Further improvement may be obtained with rotating wheels.
4.1 Tangential blowing
With tangential blowing, air is blown in the direction of the wind through a narrow slot in
the wind tunnel test section floor. The blowing device consists of a slot adjustment
mechanism and a tubular settling chamber. The principle is that so much air is added to the
boundary layer that the momentum deficit in the boundary layer of the wind tunnel floor is
reduced to zero. This can be reached exactly at only one downstream distance from the
blowing slot.
The system of the LLF is located 4.5 m upstream off the balance centre. The slot spans a
length of 6 m and has a variable width between 0 and 5 mm. In order to arrive at a
homogeneous spanwise velocity distribution at the exit of the slot, the settling chamber is
divided into six individually controlled lateral sections, each provided with a porous
smoothing plate. Pressure and temperature are monitored at the centre of the chamber.
The momentum of the thin layer of blown air must be balanced with the momentum of the
airflow in the tunnel. From calibration of the blowing device the required ratio V
j
/V

0
of the
jet velocity V
j
at the slot exit to the free stream velocity V
0
is determined as function of the
V
0
.
Figure 9 shows the setup and calibration results of the tangential blowing system. With
tangential blowing good results can be obtained, even for small ground clearances.
However, some deformation in the velocity profile remains and the effects from rotating
wheels are still ignored.
Automotive Testing in the German-Dutch Wind Tunnels

23

Fig. 9. Tangential blowing system with setup (left) and calibration results (right)
4.2 Moving belt ground plane
The most appropriate way to simulate the full-scale conditions is with a moving belt ground
plane. The momentum loss in the boundary layer is almost completely absent over the
complete length of the rolling floor and the rotation of the wheels can be enabled as well.
However, the disadvantage is that the suspension of the vehicle in the test section is no
longer simple.
The moving belt at the LLF has a maximum width of 6.3 m and a flow exposed length
between the two main rollers of 7.6 m. Two variable-speed drives of 85 kW each give the
belt a maximum speed of 40 m/s. The belt is tensioned and tracked by means of a third
roller. Its flatness is monitored during testing by a video camera and with the aid of a laser
beam on the wind tunnel side wall. Even under most severe conditions when the vehicle

exerts lift, the belt must remain very flat. To reduce friction, pressurized air is fed between
belt and support plate. The upstream ground floor boundary layer is scooped off by raising
the whole assembly 200 mm above the tunnel floor. The extracted air re-injects
automatically at the rear of the belt assembly and through the test section breathers.
The influence of the scoop and belt motion on the boundary layer properties has been
calibrated by measuring the velocity profiles at the front and the rear of the belt. Beside the
removal of the boundary layer it is essential that the static pressure in longitudinal direction
remains constant over the length of a full size passenger car to avoid buoyancy effects in the
drag data. This is effectively controlled by adjusting the rearward flaps of the belt. Figure 10
shows the setup and some flow characteristics of the moving belt ground plane at the LLF.
4.3 Rotating wheels
The effects of rotating wheels on the measured aerodynamic forces can be investigated in
combination with a moving belt ground plane that drives the wheels by friction. The car is
New Trends and Developments in Automotive Industry

24
mounted to the available sting and forces are measured with an internal balance between
sting and car.
Shock absorbers and springs of each wheel are replaced by a dual-action pneumatic cylinder
to counterbalance the wheel's weight but still be able to create an effective downward force.
Rolling resistance of the wheels is determined from the internal balance measurement
without wind.
In general, the drag is reduced by wheel rotation, but the magnitude of the drag reduction
depends strongly on the type of car and on the ground clearance of the vehicle. The closer to
the ground, the more pronounced the effect of the rolling wheels will be; see figure 8.


Fig. 10. Setup (left) and velocity profiles (right) of the moving belt ground plane
5. Acoustic measurements
The classical type of acoustic measurements with trucks in wind tunnels is based on the

measurement of the noise inside the cabin, as induced by the airflow around the truck and
measured with a small number of microphones or with a so-called acoustic head. These
cabin noise measurements can be used for the assessment of the acoustic comfort for the
truck driver.
The transfer mechanism of the noise from outside the cabin towards cabin interior is often
very difficult or impossible to determine. Typical exterior structures are mirrors, the
sunscreen above the front window, wind shields, antennas and various spoilers.
Instead of measuring interior cabin noise levels at various exterior configurations, it is more
straightforward to measure the exterior sound production. This can be realized with an
acoustic mirror or with an array of a large number of microphones.
In an acoustic mirror system a single microphone is mounted in the focal point of a
parabolic or elliptic acoustic mirror. Single-microphone measurements give overall noise
levels and do not distinct between different noise sources. In a phased microphone system
the location and strength of different noise sources can be measured by a phased array
technique, whereby on software level the time series of the microphones are analyzed.
Similar techniques are applied in radar technology and ultrasonic imaging. A description of
Automotive Testing in the German-Dutch Wind Tunnels

25
array signal processing is given by Johnson and Dudgeon (1993). A description of
applications in a wind tunnel environment is given by Underbrink and Dougherty (1996),
Piet and Elias (1997), Sijtsma (1997), Dougherty (1997) and Sijtsma and Holthusen (1999).
Figure 11 illustrates the principles of the acoustic mirror and the phased microphone array.

focal pointmicrophone
sound rays
scan plane
elliptic mirror

Phased microphone array: principle

scan planemicrophones
t
Advantage: scanning after measurements (electronically)
p
t
p
2: delay&sum
t
p
t
p
X
X
3: source plot
1: time signals

Fig. 11. Acoustic source localization measurement techniques, mirror (left) and acoustic
array (right)
Microphone arrays or acoustic mirrors have become popular in wind tunnel measurements
as a tool to locate sound sources. Microphone arrays have the advantage over acoustic
mirrors of a higher measurement speed. Mirrors have to scan the whole test object point by
point, while microphone arrays only need a short time to record the signals from which the
aero-acoustic characteristics in a measuring plane can be determined. The process of
scanning through possible source locations is performed afterwards by appropriate software
running on powerful computer hardware.
An additional advantage of a microphone array is the application inside the flow or in the
wall of a closed test section. These in-flow measurements with microphone arrays are
possible, when the self-noise of the array microphones, caused by the turbulent boundary
layer above the array, is sufficiently suppressed. With a mirror, in-flow measurements are
practically speaking impossible.

Fig. 12. Test setup examples: full-scale wing in the LLF (left) and scaled truck model in the
LST (right)
New Trends and Developments in Automotive Industry

26
In a typical test an array of 1 m diameter, containing about 140 sparsely distributed
microphones, may be mounted in or on the wall of the test section.
The microphone array technique can be successfully applied in full-scale tests as well as
model tests; see figure 12 for some examples.
The array processing delivers as its main result so-called noise maps. Figure 13 presents
some results for a scaled truck and for a full-scale truck. The two dimensional contour maps
show the distribution of noise sources in a scanned area near the truck. The noise levels are
represented by different colors. The noise maps deliver the location, frequency
characteristics and relative strength of the noise source. Additionally the array processing
delivers power spectra and overall power levels by integration over the scan area. Several of
such scan areas can be defined and processed. One scan area could cover the whole model
and other areas could only cover small details, like an outside car mirror, to allow detailed
comparisons between different model configurations.


Fig. 13. Microphone array tests results for a scaled truck model (left) and full-scale truck
with projected noise map (right)
6. Flow field measurements with a traversing rake of five-hole probes
Quantitative flow field measurements can be executed by means of a traversing rake of
multiple five-hole probes. Each five-hole probe can measure the local 3-D wind velocity
vector. At each position of the rake the wind speed vector is measured at all probe positions.
At DNW there are 18 probes at 15 mm stitch; so each time the data are read out information
on a line of 255 mm length are gathered. The rake is normally mounted vertically and

connected to a traversing mechanism which is moving at such a low speed that the local
flow field is not affected. By repeating the readings during the scan after say every 7.5 mm
displacement of the rake and repeating the scan at a vertical displacement of the rake of also
7.5 mm, a block of measuring points is filled with a horizontal and a vertical stitch of 7.5
mm. This is enough to observe flow phenomena on a rather small scale. Software tools may
provide additional information, like the strength of the vorticity in the flow. A single scan of
about 1 meter at a low traversing speed requires a measuring time of about 10 minutes.
Automotive Testing in the German-Dutch Wind Tunnels

27
The test technique is providing very nice results in as well a quantitative as a qualitative
way. Only close to the surface of the test object the flow may become disturbed by the
presence of the rake body close to the object.
An example of the setup in the LST wind tunnel is shown in figure 14, together with some
test results behind the wing tip of an aircraft model.


Fig. 14. Test setup in the LST (left) and test results behind the wing tip of an aircraft model (right)
7. Flow field measurements with PIV
The flow field in the vicinity of a test object can be measured by means of Particle Image
Velocimetry (PIV); see figure 15.


Fig. 15. Experimental setup for PIV
During PIV measurements the flow is seeded with small particles with a diameter in the
order of 10 to 100 micrometers (a kind of a light smoke). With a laser two flashing light
New Trends and Developments in Automotive Industry

28
planes are created shortly after each other, whereby the light is reflected by the particles.

The images are analyzed with a software algorithm, identifying the location of the separate
particles during the two images. Once this displacement is established, the corresponding
flow speeds in the laser light plane can be calculated. From these wind vector data other
characteristic parameters can be calculated, like the vorticity.
The technique and application for a wind tunnel environment became to growth in the last
decade of the 20
th
century. Various authors gave a general description of the principles and
possible application at aerodynamic research in wind tunnels, like Willert and Gharib
(1991), Adrian (1991), Hinsch (1993), Willert et al. (1996), Willert (1997), Kähler et al. (1998),
Raffel et al. (1998), Ronneberger et al. (1998) and Kompenhans et al. (1999).
Figure 16 shows a setup as applied for a wind turbine. A stereoscopic set-up of the cameras
enables the determination of the three dimensional flow field characteristics. One camera
was directed from above to the horizontal light sheet, the second camera was looking from
underneath. This configuration was fixed and could be moved as a whole from one location
to another. This fixed set-up of cameras and light sheet allows a system calibration in
advance outside the wind tunnel and avoids time consuming re-calibration.


Fig. 16. PIV set-up on a wind turbine
PIV measurements result in vector maps of the velocities in the area where the cameras are
focused to. Figure 17 shows some test results at two different setups: underneath a military
aircraft and behind the tip of a wind turbine rotor.
The application of PIV in large wind tunnels gives some specific challenges:
• large observation areas requested,
• large observation distances exist between camera and light sheet,
• much time needed for the setup of the PIV system,
• strict safety measures required for laser and seeding,
• high operational costs of the wind tunnel.
In spite of these stringent requirements, the PIV technique is very attractive in modern

aerodynamic research. It helps in understanding unsteady flow phenomena such as shear
Automotive Testing in the German-Dutch Wind Tunnels

29
and boundary layers, wake vortices and separated flows. PIV enables spatially resolved
measurements of the instantaneous velocity field within a very short time and allows the
detection of large and small scale spatial structures in the flow. The PIV method can further
provide the experimental data necessary to the validation of an increasing number of high
quality numerical flow simulations.


Fig. 17. PIV measurements: exhaust flow of a fighter engine (left) and tip vortex behind a
wind turbine rotor (right)
8. Deformation measurements
Within wind tunnel investigations model deformation measurements are possible with
techniques like Projection Moiré Interferometry (PMI), Projected Grid Method (PGM) and
Stereo Pattern Recognition (SPR) system.
SPR is an optical, non-intrusive method and requires a stereo setup of cameras; it is based on
a three-dimensional reconstruction of visible marker locations by using stereo images.
Stereo imaging and 3D-reconstruction can be used to determine object locations and their
motion with time. There are two possible approaches. If the positions of two (or more)
cameras and their optical characteristics are known exactly, a three-dimensional
reconstruction is very straightforward. From two images of a certain marker on the object in
three-dimensional space by two cameras, the location of this marker can be determined by
regarding the images as a result of certain translations and rotations and a final projection
on the camera image plane. After calculating transformation matrices for both cameras,
derived from the exact set up of the camera positions, the transformation equations for each
marker image can be constructed, resulting in an equation system that can be solved by a
"least square"-method. A disadvantage of this direct method is that normally the camera
positions are not known very exactly. Especially the direction of the optical axis of the

cameras and the rotation about this axis can only be measured approximately. In this case a
different approach can be used; see figure 18.
New Trends and Developments in Automotive Industry

30

Fig. 18. Setup for SPR with six markers (left) and car roof deformation measurement results
from PGM (right)
The transformation matrices can be calculated if the locations of at least 6 markers are
known exactly in 3-dimensional space and their images can be detected in both camera
views.
Tests have shown that an accuracy of 0.01% of the complete object space can easily be
obtained. Measurement accuracy is better than 0.4 millimetres.
9. Flow visualisation techniques
A commonly used flow visualization instrument is a hand-held smoke rod. Oil is ejected
through a heated, small tube, whereby the oil is evaporated. Using non-coloured oil results
in white smoke, that follows the major flow streamlines and fills wakes and separation
zones with smoke.


Fig. 19. Laser stroboscope technique; setup with rotating mirror (left) and frozen flow field
showing vortices (right)
In combination with a laser light sheet the flow structure is made visible within that plane.
A laser light sheet can be created when a laser beam is diverged through a circular cylindrical
lens. It is also possible to reflect the laser beam on a rotating mirror. The continuously rotating
Automotive Testing in the German-Dutch Wind Tunnels

31
laser beam gives the same effect as a laser light sheet, provided that the rotational speed is
high enough. By varying the rpm value of the mirror, stroboscopic effects are achieved and

periodic flow phenomena can be analyzed. Figure 19 shows a sketch of the setup of the laser
light sheet by means of a rotating mirror and an example of vortices visualized with this
technique.
Other techniques to visualize the flow are using tufts or oil on the surface of the test object.
Tufts are small filaments of cotton or plastic, which are mounted on the surface with magic
tape or alike. Tufts follow the local streamlines along the body or behave like small waving
flags in separated flow regions. Depending on the material, the tufts may reflect ultraviolet
light. Tufts are easy to mount and provide useful basic information. They can be used in a
continuous way when the flow direction is changed.
Another technique to visualize the flow is by using a kind of oil on the surface. Depending
on the applied oil, transition zones from laminar to turbulent flow may become visible or
the separation zones of the flow. Certain oils also reflect ultraviolet light, enhancing the
pictures.
Disadvantages of using oil are among others the contamination of the wind tunnel and the
test time needed to establish a well-developed oil pattern.
10. Wind tunnel blockage corrections
Testing vehicles in a wind tunnel introduces disturbing effects from the finite dimensions of
the airflow. In case of a ¾ open test section the flow from the exit nozzle may have some
divergence, leading to a streamline divergence near the vehicle which is somewhat larger
than in the unconfined real condition. This results in too low wind loads. In case of a closed
test section the streamline divergence near the vehicle is reduced because of confinement by
the wind tunnel walls. This results in an increase of the kinetic pressure at the tested object
and thus an increase of the measured wind loads.
Corrections are needed, especially for closed test sections and relative large vehicles
compared to the wind tunnel cross section dimensions.
In case of a closed test section it is possible to correct by measuring wall pressures in the
vicinity of the vehicle. This is however rather elaborate and not common practice. More
usual is to correct the reference kinetic pressure analytically, e.g. by a formula that
incorporates the measured drag. This latter correction method is basically a base-pressure
correction method that started with the work of Maskell (1963), Gould (1969) and Awbi

(1978). An empirical blockage correction for trucks in the LLF wind tunnel is described by
Willemsen and Mercker (1983). A description of a blockage correction method for
automotive testing in a wind tunnel with closed test section is described by among others
Mercker (1986).
11. References
Adrian, R. J. (1991), Particle-imaging techniques for experimental fluid mechanics, Annual
Reviews Fluid Mechanics, Vol. 23, pp. 261-304.
Awbi, H.B. (1978), Wind tunnel wall constraint on two-dimensional rectangular section
prisms.
Dougherty, R.P. (1997), Source location with sparse acoustic arrays; interference
cancellation, presented at the First CEAS-ASC Workshop: Wind Tunnel Testing in
Aeroacoustics, Marknesse.
New Trends and Developments in Automotive Industry

32
Gould, R.W.F. (1969), With blockage corrections in a closed wind tunnel for one or two wall-
mounted models subject to separated flow, Aeronautical Research Council Reports and
Memoranda, no. 3649.
Hinsch, K.D. (1993), Particle image velocimetry, Speckle Metrology, Ed. R.S. Sirohi, pp. 235-
323, Marcel Dekker, New York.
Johnson, D.H., Dudgeon, D.E. (1993), Array Signal Processing, Prentice Hall.
Kähler, C.J., Adrian, R.J., Willert, C.E. (1998), Turbulent boundary layer investigations with
conventional- and stereoscopic particle image velocimetry, Proceedings 9th
International Symposium on Application of Laser Techniques to Fluid Mechanics, Lisbon,
paper 11.1.
Kompenhans, J., Raffel, M., Dieterle, L., Dewhirst, T., Vollmers, H., Ehrenfried, K.,
Willert, C., Pengel, K., Kähler, C., Schröder, A. and Ronneberger, O. (1999), Particle
image velocimetry in aerodynamics: technology and applications in wind tunnels,
Journal of Visualization, Vol. 2.
Maskell, E.C. (1963), A theory of the blockage effects on bluff bodies and stalled wings in a

closed wind tunnel, Aeronautical Research Council Reports and Memoranda, no. 3400.
Mercker, E., Knape, H.W. (1989), Ground simulation with moving belt and tangential
blowing for full-scale automotive testing in a wind tunnel, SAE Paper 890367,
Detroit.
Mercker, E., Wiedemann, J. (1990), Comparison of different ground simulation techniques
for use in automotive wind tunnels, SAE Paper 900321, Detroit.
Mercker, E. (1986): A blockage correction for automotive testing in a wind tunnel with
closed test section, Journal of Wind Engineering and Industrial Aerodynamics, 22.
Piet, J.F., Elias, G. (1997), Airframe noise source localization using a microphone array, AIAA
Paper 97-1643.
Raffel, M., Willert, C., Kompenhans, J. (1998), Particle image velocimetry - a practical guide,
Springer Verlag, Berlin.
Ronneberger, O., Raffel, M., Kompenhans, J. (1998), Advanced evaluation algorithms for
standard and dual plane particle image velocimetry, Proceedings 9th International
Symposium on Application of Laser Techniques to Fluid Mechanics, paper 10.1, Lisbon.
Sijtsma, P., Holthusen, H. (1999), Source location by phased array measurements in closed
wind tunnel test sections, NLR-TP-99108.
Sijtsma, P. (1997), Optimum arrangements in a planar microphone array, presented at the
First CEAS- ASC Workshop: Wind Tunnel Testing in Aeroacoustics, Marknesse.
Underbrink, J.R.; Dougherty, R.P. (1996), Array design of non-intrusive measurement of
noise sources, Noise-Conference 96, Seattle, Washington.
Willemsen, E., Mercker, E. (1983), Empirical blockage corrections for full-scale automotive
testing on straight trucks in a wind tunnel, NLR TR 83065 L.
Willert, C., Raffel, M., Kompenhans, J., Stasicki, B., Kähler, C. (1996), Recent applications of
particle image velocimetry in aerodynamic research, Flow Measurement and
Instrumentation, Vol. 7, pp. 247 -256.
Willert, C. (1997), Stereoscopic digital particle image velocimetry for application in wind
tunnel flows, Measurement, Science and Technique, Vol. 8, No. 12., pp. 1465 – 1479.
Willert, C.E., Gharib, M. (1991), Digital particle image velocimetry, Experiments in Fluids,
Vol. 10, pp. 181-183.

3
Monitoring and Fault Diagnosis in
Manufacturing Processes in the
Automotive Industry
Roberto Arnanz Gómez, María A. Gallego de Santiago, Aníbal Reñones
Domínguez, Javier Rodríguez Nieto and Sergio Saludes Rodil
CARTIF Technology Centre
Spain
1. Introduction
At present production systems in car manufacturing processes are under high demand
requirements and maintenance plans are of great importance in order to achieve the
production objectives. The main goal of the maintenance is to increase the operativity of the
plant and the machines involved in the manufacturing process, avoiding all unexpected
stops. Preventive maintenance has been the solution adopted by most factories for years.
Based on past experience or on machines suppliers specifications, the maintenance manager
decides when to check or replace the machines or some of their components to guarantee
their operation without faults until the next maintenance stop. This implies two kinds of
costs for the factory: checking a lot of equipment (time and staff costs) and replacing
components that may be in good conditions.
That is why knowing the actual state of the different parts and machines of the factory is
so important for a good management of the plant. The increasing automation of the plants
allows to acquire, store and visualize lots of variables of the process. Most factories have
nowadays SCADA systems that allow supervision of processes and equipment giving a
valuable information about them. However it is not easy to manage this great amount of
information for different reasons. First of all the sample rate of these variables usually
hides their dynamic behaviour. Also the complexity of the processes makes it difficult to
identify all the relations and dependencies between variables, so it is not possible to
determine a wrong operation looking only the variation of a few variables without taking
into account how the rest are changing. The number of variables and data acquired in the
whole factory makes it impossible for a human supervisor to process all that information,

relate it to past data and try to find out if something is going wrong. Although his
experience will allow him to detect some problems it is evident that he needs some help to
succeed in his work.
Predictive maintenance is a methodology that improves systems availability and contributes
to cost reduction and increase of useful life of production assets. It comprises different
techniques to process acquired data from the factory to determine machines state and
predict how they will work in the future. The variety of problems that must be solved makes
the design of a predictive maintenance system be a very complex task where different
New Trends and Developments in Automotive Industry

34
knowledge areas must be integrated. It is very important to know the state of the art in all of
them and sometimes introduce innovations for applying the solutions to particular cases.
Next sections explain the main components of a predictive maintenance system and how it
was implemented in real industrial problems of the automotive industry. An effort has been
made in order to choose case studies that offer a wide range of the possible techniques to
use, combining classical solutions with newer ones.
2. Structure of a fault detection system for the automotive industry
The core of any predictive maintenance system is a fault diagnosis system able to detect
failures not only when they are happening, but also a pre-failure behaviour. It is an
advanced solution for the supervision level of the factory where in most cases only SCADAs
and alarms based on variables values are considered. One of the main advantages of
predictive maintenance is its ability to provide useful information to the human supervisor
showing what the real state of a plant or machine is and helping him in the planification of
the factory operation. It is also capable of substituting the human operator in some systems
taking decisions such as stopping the operation in case of a critical fault or scheduling
maintenance operations.
The three main components for any fault diagnosis system are data acquisition, signal
processing and decision making. These three components must be designed jointly because
the requirements or outputs of one of them will affect the others. Their complexity level will

depend on the application and how the symptoms of the faults can be found.
Data acquisition is the first stage of every diagnosis system. This component consists of all
the sensors, signal transmission systems, acquisition devices and storage equipment.
Sensors are a key component of the fault detection system because they provide all the
information the system will have to deal with, although in some cases information coming
from production management systems can be useful. In some cases those sensors can be
shared with other tasks such as control or supervision and they are included in the machine
or plant during its design. But in most cases predictive maintenance is not taken into
account during the design of the machines and new sensors are usually required. This
occurs specially when predictive maintenance must to be applied to old machines because
they start to be a bottle neck in the plant due to their unexpected faults. Electric current,
voltage, accelerometers and temperature sensors are of common use for diagnosis systems.
Some applications require more specific sensors, like photodiodes and spectrometers. The
selection of the appropriate sensor and acquisition system can be determinant for the
success of the application because they must guarantee that the collected data have the
information of the state of the machine. Capture and synchronize data from sensors of
different nature and variables with different dynamics can be an interesting problem to
solve and sometimes requires specific programming or storaging methods designed ad-hoc.
In the signal processing stage, signals acquired and/or stored by the data acquisition
component are processed. This includes common signal treatment like filtering that is used
to eliminate noise. However, the most important part in signal processing is feature
extraction. Feature extraction consists in looking for a particular behaviour in the signals
that allows to identify the faulty or pre-faulty states. There are a wide variety of feature
extraction techniques and the one used depends on the problem at hand. For example, one
of the most common feature extraction techniques is the Fourier Analysis, which gives
information on the distribution of energy power associated to different frequency ranges in
Monitoring and Fault Diagnosis in Manufacturing Processes in the Automotive Industry

35
the signals acquired by the sensors. This content changes when a fault occurs or is close to

occur. Besides this, feature extraction techniques in the time domain are also useful. Some
problems require the use of very specific feature extraction techniques, like the estimation of
electron temperature. The final stage is the decision making where the features that have been
extracted from the signals generated by the sensors, have to be classified in order to determine
the state of the system. The classification is the base of the decision making process, so it has an
important role in the fault diagnosis scheme. In some cases, classification can be done by
merely checking the features values against a threshold, although selecting the threshold value
could be a hard problem to solve. In other cases, more sophisticated non-linear classifiers, like
neural networks, neuro-fuzzy systems or support vector machines have to be used. Besides
this, features time evolution is also of great importance because it allows to perform trend
analysis, which is one of the basis of the fault predictive capabilities of the fault detection
systems. The lack of historical data is the main problem that must be solved when designing
the decision making component. It can be sometimes a problem to decide what is the optimal
classification method to use, and it is always an added difficulty to fix the parameters of the
system. Usually conservative strategies are used. This leads to a great number of false alarms
during the initial phases of the predictive maintenance system implementation. Human
experts supervision and knowledge is one of the main supports for a good design of the
decision making system and its configuration.
3. Case studies
3.1 Case study 1: Multitooth machine tool
Machine tools represent one of the main examples of highly automated components
(Altintas, 2000). In spite of this automation, the cutting process has an inherent degradation
(Astakhov, 2004), which is one of the main problems to be overcome. Other aspects to
consider are workpiece tolerance deviations, ensuring a correct evacuation of the chips,
changing of worn tools and, if necessary, stopping the machine if abnormal working
conditions appear (for example chatter). So, to achieve the desired level of autonomy for this
kind of machines, it is necessary to develop the monitoring and diagnosis of the cutting
process. Many different kinds of machine tools are used in the automotive industry. Among
them, the so called Multitooth Machine Tools represent the most challenging ones, from the
diagnosis point of view, due to high number of inserts susceptible to break, and the different

machining operations integrated within the same tool. The tools analyzed in this chapter are
used in the car industry for mass production of different mechanical parts, such as the
crankshaft or the camshaft of car engines. These tools are complex ad hoc devices built with
many cutting inserts (up to 250, depending on the machine) of different kinds (roughing and
finishing) presented in Fig.1(b) and for different operations (turning, milling or broaching)
within the same tool, as shown in Fig.1(a). The configuration of the tool is based on multiple
tool holders specially designed for the particular operation of the mass production line.
Such complexity is necessary to achieve the required high metal removal rate.
3.1.1 Data acquisition in machine tool environment
Regarding the main three components of a fault detection system (data acquisition, signal
processing and decision making) an optimal selection of sensors is of paramount importance
to obtain valuable information from the environment of the machine tool that should be
correlated with the abnormalities to be detected. Different signals susceptible of

New Trends and Developments in Automotive Industry

36

(a) Layout of multitooth tools used in the car
industry
(b) Different inserts in the multitooth tool
Fig. 1. Multitooth tools used in the car industry
having correlation with tool wear and the breakage of inserts in the multitooth tool, are
shown in Fig.2. Among others the following are the most common in the literature:
Noise: can be measured in the environment of the tool using microphones (Fig. 2(a)). Al-
though noise gathers information coming from the whole machine tool environment,
this measure can be very valuable for the first analysis of the machining cycle through
the analysis of the time-frequency representation like the spectrogram.
Vibration: measured with accelerometers in one of the main shafts of the machine tool (Fig.
2(b)). As the wear increases in the tool an abnormal increase in the vibration also

occurs and can often lead to bad surface quality.
Temperature: the increase of the tool wear causes an increase in the temperature due to an
excessive friction. Using sensors like pyrometers, the temperature of the machined
surface can be easily measured after the machining has been completed (Fig. 2(c)).
Electrical power consumption: can be measured from the output signals of the frequency
converters (for the usual case of AC drives) for every motor that moves the multitooth
tool and moves the workpiece (usually rotation movement). Fig. 2(d) depicts the
example of rms electrical power consumption of the two drives of an example tool:
feed and rotation of the tool holder. These kind of signals show clearly the different
parts of the cycle and the grouped attack of the inserts in the tool.
In order to analyze the sensitivity of every recorded signal, the measurements have to be
done over the useful life of several consecutive tools. After that, every set of signals is
statistically analyzed to extract global information for comparison and to decide whether
there is a correlation with the degradation of the tool, or other abnormalities that could have
been recorded. In (Reñones, Rodríguez & Miguel, 2009) are presented the results of such
analysis that lead to choose the electrical power consumption as the most appropriate signal
for use in the diagnosis of the multitooth tool. This signal showed the best signal-to-noise
ratio for the evolution of the wear and was the most cost-effective measure: non-invasive,
moderate sensor cost (inexpensive if appropriate signals are available at the drive converter)
and high reliability of the measure in comparison with other measures like noise and
vibration, because of the high influence of the sensor location. Fig. 3 shows the evolution in
electrical power consumption in a particular zone of the analyzed tool. It is clear the increase
of the power due to the wear and the abrupt decrease after the tool reaches its useful life and
it is changed by a new one.
Monitoring and Fault Diagnosis in Manufacturing Processes in the Automotive Industry

37
-0.5
-0.3
-0.1

0.1
0.3
0.5
010203040
Time (s)
Sound pressure (u.a.)
(a) Sound pressure recorded during the
machining of a car crankshaft supports
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
0.2
0 10203040
Time (s)
Vibration (g)
(b) Vibration amplitude of the main rotation tool
axis recorded during the machining of a
car crankshaft supports
0.41
0.42
0.43
0.44
0.45
0.46
0.47

0.48
0.49
0.5
0 8 16 24 32 40 48 56 64
Time (s)
Temperature (u.a.)
Temperature of
the machined
surface
Temperature
increase
Ambient
temperature
(c) Temperature support scanned after the
machining cycle
-30
-20
-10
0
10
20
30
0 102030405060
Time (s)
Current (A)
Feed
Rotation
Roughing
Finishing
(d) Electrical power consumption of the feed and

rotation electrical drives

Fig. 2. Signal more common in the literature


Fig. 3. Evolution of the tool wear using the electrical power consumption.
3.1.2 Signal processing of the electrical power consumption
Once the electrical power consumption has been chosen as the desired signal for diagnosis
of the tool, it is time to extract the part of the electrical power consumption that belongs to
each insert or every group of inserts that attacks the workpiece simultaneously. This process
is known as signal segmentation and can be formulated as the automatic decomposition of a
signal into stationary or transient pieces with a length adapted to the local properties of the
signal (Basseville & Nikiforov, 1993).
Firstly, the number of segments that must be extracted has to be defined, taking into account
different aspects of the machining process, such as the different kinds of cutting inserts, the
workpiece material, changes in the cutting conditions, changes in the PLC programming,
different mechanized zones of the workpiece and the layout of the tool.
Among the different alternatives for making the segmentation of the electrical power
consumption signals, the use of auxiliary signals not directly affected by the machining,
such as, for example, the sampled speed reference of the machining cycle, or its acceleration,
ensures a reliable segmentation avoiding false alarms in the detection of a fault in the tool.
Once the different signal segments are extracted, the next step is to obtain the model for
every segment that should be sensitive to electrical power consumption changes caused by a
New Trends and Developments in Automotive Industry

38
fault in the tool. Another goal of this step is to reduce the amount of information used in the
following steps of the diagnosis scheme. There are different methods available to make such
reduction (Reñones, Rodríguez & Miguel, 2009). Among them, the calculation of statistical
parameters is a straightforward reduction of information. Only those which presented a

greater sensitivity to the variations produced by the failures in the tool must be chosen in
order to reduce the amount of data for detection of failures in each group of the tool. With
the appropriate statistical parameters chosen, the change detection problem can be stated as
the detection of a change in a set of random variables. The change detection is usually
carried out using a so called stopping rule, as presented in (1); that is, a function of the
random variables y
k
that exceed a preset threshold λ in case of abrupt change. The
parameter t
a
represents the estimated time of change at which the stopping rule is true for
the first time (Basseville & Nikiforov, 1993).

{
}
1
inf : ( , , )
ann
tngyy
λ
=…≥ (1)
This problem is frequently solved from a statistical point of view. In Fig.3 an example of
abrupt change that must be detected can be seen. The following requirements must be taken
into account to solve this change detection problem:
• The segmentations or electrical power consumption trends are non-stationary, so an
adaptive detection scheme is needed.
• The changes must be reliably detected, and the false alarms due to occasional electrical
power consumption changes must be avoided.
• A mean time between false alarms (MTFA) must be fixed.
• The change detection must be fast enough to avoid serious damage to the whole tool

and machine.
• The changes can be abrupt decreases (in case of breakage) but also abrupt increases due
to the loss of an insert or an abnormal wear rate caused by the breakage of previous
inserts.
Among the different alternatives that can be used to detect abrupt changes (Reñones,
Miguel & Perán, 2009), the algorithm based on an adaptive local linear model of electrical
power consumption showed the best performance in terms of reliability, and an extremely
low computational cost. The algorithm is based on the detection of linear regression outliers.
In the present case, the outliers are recorded points with an electrical power consumption
out of normal variation due to a breakage (abrupt decrease) or abnormal wear rate (abrupt
increase).
Due to the fact that the evolution of the electrical power consumption trends are not linear
as the wear increases, this detection scheme must be implemented using a moving data
window, let’s say of size L.
The outlier detection algorithm is done through the calculation of statistical parameter t
i
de-
fined in (2). This statistical parameter follows a Student’s t-distribution. Under no fault in
the tool and hence no change in the electrical power consumption, the residuals t
i
should
remain in the interval ±t(1 –
α
/(2L), L – 3) of confidence
α
. These bounds of the interval are
also known as the critical level or threshold.

()
ˆ

1
i
i
Ri ii
e
t
Sv
=
−
(2)
Monitoring and Fault Diagnosis in Manufacturing Processes in the Automotive Industry

39
In order to adjust the algorithm in an optimal way, some performance measures must be done
and it must be taken into account the variation range of the different parameters for the
algorithm (window size L and the critical level or threshold). To make the detection robust, an
additional parameter can be added, such as the amount of consecutive detected outliers.
In quality control this is called a run test. In fact, this robust mechanism is not particular to
this detection scheme and can be applied to other detection algorithms.
In order to optimally adjust the parameters of the detection change algorithm, performance
measures must be done, such as (Gusstafson, 2000): MTFA (Mean Time between False
Alarms), MTD (Mean Time to Detection) and, MDR (Missed Detection Rate).
The optimal algorithm adjustment is performed by fixing either the performance measure
MTFA or MTD, and the parameters of the algorithm are chosen to minimize the other
performance measures. The presented algorithm have been evaluated with data coming
from the machining of more than 30000 workpieces. As the exploration of the whole range
of parameters for the change detection algorithm is unapproachable, some restrictions and
assumptions were added to cope with the problem. For the window size L, it seems
reasonable to choose a value lower than the mean time between faults. For the test set used,
it is approximately 300 workpieces, then the interval for this parameter was set as [40,100]

workpieces.
The run test, represented as R, influences the speed of detection. After studying historical
data and taking into account the protection of the tool, an interval of [2,6] workpieces seems
reasonable. The threshold interval was [2,7] and for the residuals was fixed as an interval
with a confidence level from 0.1 to 0.001. Two tests have been done to study the relationship
between the different parameters, where the threshold is varied in the preset interval and
the other two parameters are fixed at the midpoint of their own interval.
In Fig. 4 is presented an example of such performance measures. Detailed analysis of these
graphics can be found in (Reñones, Miguel & Perán, 2009). It is straightforward to see that
an increase in the threshold (horizontal axis of the graphics) leads to a more reliable
detection (higher MTFA) but fewer faults are detected as shown in the third row of graphics.
This exploration of parameters influence let to finally make an optimal adjustment in the
parameters for the change detection in the different electrical power consumption trends for
the different zones of the multitooth tool.
The result of this step is a list of thresholds for every zone of the tool. Positive thresholds can
be adjusted to detect abrupt increases of the electrical power consumption due to an
abnormal wear rate (called as overload), and also abrupt decreases due to a breakage of one
or more inserts in the tool.
3.1.3 Decision making process for the machine tool diagnosis
The last step of the methodology used to detect faults in the multitooth tool is the so-called
Decision-making process as presented in section 2. In this step, using the information coming
from the change detection algorithm and other information of the state of the system, an
effective declaration of the fault in some zone in the tool is done. That means, for example,
that the machine tool will be stopped at the end of the current cutting cycle, and the
operator will fix the problem based on the information of the diagnosis system: the faulty
zone of the tool and the type of fault (overload or breakage).
The electrical power consumption signal gives the best signal-to-noise ratio to detect faults,
as was presented in section 3.1.1. On the other hand, this signal exhibits sensitivity (abrupt
changes in the signal) to other phenomena that may cause false alarms which must

New Trends and Developments in Automotive Industry

40
10
100
1000
10000
100000
234567
Umbral
MTFA
R=2
R=6
10
100
1000
10000
100000
234567
Umbral
MTFA
1
2
3
4
5
6
7
8
234567

Umbral
MTD
R=2
R=6
1
2
3
4
5
6
7
8
234567
Umbral
MTD
L=40
L=100
@L=70
@L=70
@R=4
@R=4
50
60
70
80
90
100
234567
Threshold
% Detected

@L=70
R=2
R=6
50
60
70
80
90
100
234567
Threshold
% Detected
L=40
L=100
@R=4

Fig. 4. Performance measures for the linear regression change detection algorithm with
parameters: L ∈ [40,100], R ∈ [2,6] and λ ∈ [2,7].
be taken into account, such as a tool changed by a new one, changes in the material of the
workpieces (foundry or steel), compensation adjustments in the inserts made by the
operators to achieve the desired tolerances, the warm up process after a long stop, etc. To
prevent false alarms caused by any of these events, it is necessary to protect or disable the
change detection algorithm. Protective measures that can be taken to avoid false alarms are
to use output signals from the PLC governing the machine tool (new tool, material change,
etc), or to inhibit change detection when changes affect the whole recorded signal or there
are sample points separated too much time.
3.2 Case study 2: Car painting cabinet
This case study shows a predictive maintenance system currently operating in an assembly
car factory, specifically in painting cabinets section. It has been working for thirteen years
now and serves as a valuable tool for anticipating to breakdowns all along the plant,

optimizing equipment performance and reducing unplanned shutdowns and incidents. This
predictive maintenance system is based on mechanical vibrations analysis techniques
applied on the motor-fan sets operating in painting cabinets.
The predictive maintenance for this kind of installations can be performed in two ways.
With online analysis systems or with hand-held, walk-around vibration analyzers. For
Monitoring and Fault Diagnosis in Manufacturing Processes in the Automotive Industry

41
extremely large operations and/or very expensive equipment, the first approach is the most
cost effective and has repeatedly shown to saving money.
The main advantage of an on-line dynamic vibration monitoring system is that the data
acquisition is made continuously. This allows to check past values and to know the
evolution of the state of the machine, providing a more reliable diagnosis that off-line data
acquisition systems cannot offer. Most of the on-line systems use some kind of acquisition
system architecture that involves input channels multiplexing many vibration sensors. This
results in a scan rate that varies according to the system scheduler. Another advantage to an
on-line dynamic vibration monitoring system is that there is no labour cost to acquiring the
data and minimal labour cost for identifying machine faults.
The disadvantages of these systems are that they are the more costly systems to implement and
maintain as they include maintaining a full time vibration analyst, and installing a wired
network to get the signals from the sensor to the analysis system. Furthermore, the software and
hardware that make up the system typically require an extensive maintenance contract as well.
Hand-held, walk around vibration analyzers only provide trending information to identify
that a potential problem exists, and do not provide the detailed information necessary to
determine the cause of the problem. The supervision is done only at specific moments and it
does not provide a trend of vibration levels. Moreover, they require skilled vibration
analysts to interpret the data and, without continuous monitoring, problems in between
rounds could be costly.
3.2.1 Problem description
The plant under study consists of a series of motor–fans that keep painting cabinets under

very strict temperature and humid conditions. In some cases air must be put into these
cabinets and in some others air is taken out of them. The target is to keep working
atmosphere under control in such a way that safety and sanitary conditions are guaranteed
for the staff. Moreover, in order to achieve a good production quality, it is required that air
inside the cabinets is at the right temperature, filtered and keeping an adequate relative
humidity that prevents varnish thinners from evaporation. It is also necessary to extract the
air from the cabinet, in order to eliminate polluting elements.
For each motor-fan the fan is driven by an electric drive whose rotation movement is
transmitted to the fan through a couple of pulleys, one attached to the fan and the other one
to the drive, together with a belt. Both the electric drive and fan are mounted on an elastic
structure that keeps the set isolated from the high frequency excitations of the structure and
at the same time, this base structure is not affected by the mechanical vibrations coming
from the electric drive and fan.
This assembly plant is able to produce around 1.200 cars every day along three shifts,
depending on demand needs. To achieve this, it is mandatory to ensure that every machine
is working under optimal conditions avoiding unexpected breakdowns which could lead to
stops and subsequent lost of production. Therefore, a predictive maintenance system is
needed. A thorough analysis of the related machines has led us to consider the following
sources of mechanical vibration that could be the cause of potential failures:
1. Defect related mechanical vibrations: Unbalance, misalignment, looseness, defects in
bearings, blade breakage and defects in belts.
2. Mechanical vibrations related to natural frequencies: Natural frequencies of the base
structure, natural frequencies from any part of the machine structure and natural
frequencies from other elements outside the machine.
New Trends and Developments in Automotive Industry

42
3.2.2 Predictive maintenance system
The system consists of an industrial computer in charge of data acquisition, communication
protocols and the calculation of spectra and alarms (DCS station in Fig. 5) to which up to

four nodes are connected through a LAN. They are multiplexors and receive signals from
accelerometers placed on the machines. Analysis and diagnosis tasks are carried out by
means of a PC (MD station). This PC has a communication module that allows remote access
to the data, so that it is possible to perform the same tasks from a remote computer, outside
the factory. Fig. 5 shows the layout just described.


Fig. 5. Predictive maintenance system schema
In each motor-fan two accelerometers have been placed to register mechanical vibrations from
the electric drive and the fan, which is the most sensitive part to be monitored in this case. The
related bandwidth is 20 kHz, which is enough for the application under study. They have been
placed in radial position, as close as possible to the bearings near the pulleys.
The signals from the accelerometers reach one of the four multiplexors (nodes from Fig. 5)
inside which they are displayed along 32 channels, and finally get to the industrial computer
where they are registered and sent to the PC for further analysis. As soon as an abnormal
value is detected, an alarm shows up on the screen so that subsequent actions can be taken
in order to solve the problem arising. This scheme is the same for every motor-fan being
monitored.
The system is automatically registering data on a daily basis. At the same time, mechanical
vibration levels, process variables and alarm levels are being checked for the plant.
It is possible to register three kinds of data: gross scan, spectrum and time signal:
Gross Scan: These data constitute a unique signal taken from a DC stationary signal o
calculated from an AC dynamic signal, as for example, a RMS one.
The gross scan measurement from each sensor is compared to a reference value that serves
as an alert. After this, the measurement is used to update the related maximum and
minimum values that will be finally registered in the database. Whenever any gross scan
measurement exceeds the alert value, it is first registered in the database, then it is updated
for the DCS, and finally, the related spectrum and time signal are recorded as alarm related
data for the specific sensor.
Monitoring and Fault Diagnosis in Manufacturing Processes in the Automotive Industry

43
Spectrum: They are calculated from a related time signal and further processed in order to
get specific information at certain frequencies associated with the potential defects of the
machine being monitored. In this case, this is done through what we call Analysis
Parameters Set (APS). Gross scan data are registered for every signal once per each data
registered cycle. Spectrum data are recorded for one or two signals once per each data
acquisition cycle and time signals are registered simultaneously to the spectra.
Time signal: This software allows the user to visualise data in the time domain, which can
be very useful is specific situations, though no further analysis is being performed in the
case under study.
The system allows to define up to 12 parameters directly related to the frequencies or
selected ranges of them that are of interest in order to characterize (detect) certain types of
machine failures. There are many possibilities to choose different types of Analysis
Parameter Sets and the most widely used are briefly described next:
Total Energy: This value represents all of the energy of a signal. Because of the nature of the
FFT, the first two points of the spectrum are excluded from this summation.
Energy within a Frequency Range: The energy between the two specified points of a
spectrum will be summed.
Non-Synchronous Energy within a Frequency Range: The energy between the two
specified points, which is not an integer multiple of turning speed, is summed.
Synchronous Energy within a Frequency Range: The energy between the two specified
points, which is an integer multiple of turning speed, is summed.
Synchronous Peak: The signal is synchronously sampled to determine the energy at
harmonic of running speed. In order to use this parameter, the sensor must have a
tach pulse defined for it.
HFD (5k-20kHz) High Frequency Detection: An additional collection of vibration data is
made from which the energy from 5.000 Hz to 20.000 Hz is computed. HFD
sometimes useful in detecting bearing faults at an early stage.
RPM: This field displays the RPM for this sensor. A good way to ensure that the tachometer

definition is set up correctly and the RPM ratio is correct is to compare the reported
RPM with the expected value.
3.2.3 Practical example
Next, an example of a fault is showed. In this case, a progressive defect in the fan bearings
has been detected. As soon as the pre-alarm level is reached all the related parameters are
supervised, and once the system indicates the alarm level has been exceeded, the faulty
bearings are replaced. This kind of fault is best detected using the energy within a frequency
range parameter. For this kind of defect several frequency ranges have been selected in
order to assess the degree of severity of the fault. When a bearing defect is first detected (just
within a unique frequency range), the machine will still be able to work under acceptable
conditions long before it is advisable to replace the damaged bearings. Therefore, when
some ranges are affected simultaneously the fault is considered severe enough so as to
recommend the replacement of the faulty pieces. Fig. 6 shows the trend followed by
mechanical vibrations for five related consecutive frequency ranges. They all have the same
performance, giving precise information on the very moment when the failure first
appeared. Then, it became more and more important until the alarm level was reached, and
finally it can be seen the level of vibration once the faulty bearings were replaced by means
of a planned intervention, not affecting production by any means.